CoopEval - Reputation (First-Order): leaderboard

Mean payoff of an LLM agent on CoopEval (2026) under first-order reputation (random re-matching with the recent actions of co-players visible), averaged over Prisoners Dilemma, Public Goods, Travelers Dilemma and Trust Game against a population of the other tested LLM agents, each told to maximize its own points. Normalized payoff where 0 is mutual defection and 1 is full cooperation (open scale: below 0 when exploited, above 1 with side payments); higher is better.

Metric: Normalized mean payoff (open scale; below 0 when exploited, above 1 possible under side payments) under first-order reputation (random re-matching each round with co-players' own past 3 rounds visible), aggregated over four social dilemmas (Prisoner's Dilemma, Public Goods, Traveler's Dilemma, Trust Game) after shifting and rescaling payoffs so that 0 is everyone defecting and 1 is everyone playing the most cooperative action; the model's mean payoff across all cross-play match-ups with the six tested LLM agents (a uniform population), each instructed to maximize its own points, temperature 1, three repeats per combination; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3 30B A3B 2507 Instruct0.399#1100
2Claude Sonnet 4.5Low0.375#261
3GPT-4o (2024-05-13)0.344#732
4GPT-5.2Low0.325#185
5Gemini 3 FlashMedium0.284#175

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=coopeval-reputation-first-order · How It Works · Data refreshed daily, snapshot 2026-11-21.